Methods for the Intelligent Analysis of Biomedical Data

E.V. Geger1, A.G. Podvesovskii2, S.A. Kuzmin2, V.P. Tolstenok2

emiliya_geger@mail.ru|apodv@tu-bryansk.ru|wolv3333@mail.ru|tolstenok21@yandex.ru 1Bryansk Clinicodiagnostic Center, Bryansk, Russia

2Bryansk State Technical University, Bryansk, Russia

The paper discusses methodology of cleaning and analysis of small semi-structured samples of biomedical data. This methodology

is aimed at statistical evaluation of harmful production factor correlation with workers’ laboratory test data. As a result of the analysis

and interpretation of the data, a deviation from the norm is observed according to some indicators of a clinical blood test in individuals

whose occupational activity is associated with harmful factors. Conclusions are drawn about the need for further research in the group

of people whose work is related to harmful production factors. It is necessary to employ intelligent methods for analyzing possible health

risks and their negative consequences in order to make management decisions. The presented assessment methodology can be used to

create an occupational health and safety information system.

Keywords: risk assessment, data analysis, harmful working conditions, statistical methods, data cleaning, model ensembles,

Kohonen self-organizing maps.

1. Introduction

In the modern world, the necessity often arises for auxiliary

methods of preliminary disease detection at early stages. This

problem is especially characteristic of people whose labor

activity is associated with the constant impact of harmful

working conditions [6, 10].

In turn, increased level of exposure to harmful substances

and related industrial hazards significantly increase the

likelihood of developing occupational diseases and the risk of

injury [8].

Thus, occupational risk management is a complex of

organizational and technical measures which should be based on

reliable results of data analysis [9].

For the tasks of occupational risk assessment, it is necessary

to use medical data analysis methods that correspond to these

tasks. Choice of methods affects the construction of theoretical

biomedical models and characteristics of experimental studies [7,

17].

However, many experts note that biomedical data are often

unsuitable for processing using traditional software not only

because of their volume but also because of the variety of data

types and speed at which they must be analyzed [1, 3, 15].

Therefore, a system for the intelligent analysis of medical data is

required, which could aggregate and analyze heterogeneous

information coming from different sources: electronic medical

records, data from monitoring sensors, ultrasound and X-ray

apparatus and other devices [4].

If we consider real biomedical data, they have a number of

specific features that make them unstructured: presence of

various data corruptions, such as omissions, extreme values,

manual entry errors, incorrect information, high dimensionality

and heterogeneity, a large number of noisy and duplicate data.

This leads to the unsuitability of most of the data or the entire

sample for existing analysis algorithms [16].

Analysts believe that today, to solve many tasks of the

healthcare system, it is necessary to go in the direction of

structuring information and focusing on work with small samples

[5].

Particular attention should be paid to the so-called small

samples, the volume of which is about 100-200 records. This, in

turn, is very difficult and makes the use of existing methods of

data processing and analysis ineffective.

The main obstacle to the manual analysis of small samples is

the inability of the analyst to notice hidden patterns in presented

data, while special analytical algorithms detect existing patterns

with much greater efficiency. The main task of the analyst, in this

case, is to interpret the results of these algorithms and filter out

false and trivial patterns.

New analytical technologies can not only increase the

efficiency of medical institutions but also make it possible to

solve such health problems as identifying diagnoses, medical

errors, associative connection of diagnoses with results of

laboratory tests and much more.

The objective of the present research has been to assess

relationship between occupational morbidity and harmful

production factors. So the experimental sample consisted of

individuals whose occupational activities were associated with

harmful and dangerous working conditions. Another objective

was development of a new methodology focused on processing

and analyzing small semi-structured samples of biomedical data.

2. Existing Solutions

Among many proposed methods and tools for working with

small samples, it is advisable to apply ensemble data analysis

methods in combination with the basic classifier – Kohonen self-

organizing maps.

An ensemble of models is a combination of several learning

algorithms that, working together, help to build a model more

efficient and accurate than any of the models built using a

separate algorithm. That is, to find a solution for one problem or

to prove a hypothesis, not one but several models are used.

Besides, the overall operating result matters and not that of a

single separate model.

Formerly, researchers faced the problem of combining

accuracy, simplicity and ease of interpretation in one method. A

simple method could be used with a relatively easy interpretation

but coming short of accuracy or vice versa, a complex but

accurate method could be chosen that would be difficult to

interpret. Ensembles of models have become a solution to this

problem as a universal way of improving the accuracy of

methods.

Ensemble learning refers to the training of a finite set of basic

classifiers with the subsequent combination of their forecasting

results into a single forecast of an aggregated classifier. Thus, an

aggregated classifier will give a more accurate result.

The goal of combining models is to improve (enhance) the

solution provided by a separate model. It is assumed that a single

model will never be able to achieve the efficiency that the

ensemble will provide.

Self-organizing maps are a special type of artificial neural

network allowing for non-linear regression and projection of

multidimensional data onto a two-dimensional plane with

preservation of distances in their original data space [2]. This

approach can be applied in various fields, including biomedical

data processing.

Copyright © 2019 for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).

3. Proposed Theoretical Solution

To solve the research problems of working environment risk

assessment, we formed two groups:

Group I included individuals whose labor activity was

associated with exposure to harmful production factors.

Group II consisted of individuals whose professional activity

lacked harmful production factor.

The studies were carried out in the laboratory of Bryansk

Clinical Diagnostic Center; the results were reflected in the

medical information system "MAIS DC".

The list of occupational diseases was determined in

accordance with the Order of the Ministry of Health and Social

Development of the Russian Federation No. 417n dated April 27,

2012 “On the Approval of the List of Occupational Diseases”.

Those working in harmful labor conditions are at risk for diseases

associated with exposure to occupational physical factors [12].

The studies were carried out in accordance with the Order of

the Ministry of Health of the Russian Federation dated April 12,

2011 No. 302n (as amended on February 06, 2018) “On the

Approval of Lists of Harmful and (or) Hazardous Occupational

Factors and Kinds of Work that Require Mandatory Preliminary

and Periodical Medical Examinations (Surveys), and the

Procedure for Conducting Mandatory Preliminary and Periodical

Medical Examinations (Surveys)” [13].

To assess the possible risk of diseases caused by harmful

occupational factors, the values of clinical blood test scores were

used as a source of information. The experiment was carried out

in compliance with the ethical principles of biomedical research

and in accordance with the Federal Law of the Russian

Federation No. 152 "On Personal Data" [14].

For morbidity analysis, the "International Statistical

Classification of Diseases and Related Health Problems" of the

tenth revision (ICD-10) was used [11].

To solve the problem, a methodology was proposed

consisting of the following stages:

1. Data cleaning.

2. Model ensemble development.

3. Result interpretation.

To carry out the first step, a special model has been designed

and developed, the task of which is to clean the source data and

convert unstructured data into ordered ones.

The resulting model can be roughly divided into five parts.

Data import. It includes setting field names and labels,

excluding empty fields, setting data types.

Data preprocessing. It is a submodel that generates two sets

of data at the output – intact data, containing no errors, and

corrupted data, which are input to the next submodel.

Data cleaning. This submodel receives damaged data as

input and, after appropriate processing, provides output of

cleaned data that does not contain initial errors. In particular, this

submodel sets correct deviation values. In addition to the cleared

data, the output of the submodel contains data sets with incorrect

values in deviations and in scores.

Data merge. In this part, aggregation of initially complete

and cleared data takes place. In addition to this, using a code

written in the JavaScript programming language, patient’s

unique identifier is generated and all diagnoses are divided into

several records, so that three records containing one diagnosis are

obtained from one record containing three diagnoses. This, in

turn, allows for data structuring. The converted data can be used

when applying transaction analysis algorithm. Due to the errors

in the diagnosis name, it was decided to download an excel file

containing codes and names of 12257 diagnoses of the

international classification of diseases of the tenth revision [11].

In the resulting data set, the initial field “Diagnosis Name”

has been replaced by the name of the diagnosis from this file.

Data export. It creates a file with the extension .txt, suitable

for uploading to analytical platforms and further analysis.

features of this model, in addition to the fact that it does not

allow meaningless data loss, are:

1. Correction of empty values.

2. Deletion of entries containing no information.

3. Recalculation of deviations by scores.

4. Recovery of missing scores by deviations.

5. Reassignment of diagnoses by ICD codes.

6. Assigning a unique number to each patient.

7. Increasing the number of records suitable for analysis

without generating synthetic patients.

8. Download and comparison with ICD certified directory

of codes and their descriptions.

9. Uploading data that has been damaged and

subsequently corrected for additional possible analysis.

10. Converting source data into a form suitable for

intelligent analysis methods.

At the second stage, development and construction of a

model ensemble takes place. To simplify the development

process, the following algorithm must be followed:

1. Select a base model.

An ensemble can consist of classifiers of one type (e.g., only

of decision trees or only of neural networks) or of classifiers of

various types (decision trees, neural networks, regression

models, etc.).

2. Define the approach to using the learning set.

This can be resampling (several subsamples are extracted

from the original learning set, each of which is used to train one

of the ensemble models) or the use of one learning set to train all

ensemble classifiers.

3. Select a method for combining results.

Three methods are usually used: voting (the class is selected

that has been produced by a simple majority of ensemble

models), weighted voting (the result is delivered taking into

account weights set for ensemble models) and averaging (the

output of the entire ensemble is defined as the simple average

value of outputs of all models; in weighted averaging, the outputs

of all models are multiplied by the corresponding weights).

As a result of applying this algorithm, a basic ensemble of

models has been obtained (Fig. 1), which can be further modified

and complexified.

Fig. 1. Structure of the developed model ensemble

The data that has undergone preliminary cleaning at the first

stage will be the input data.

At the third stage, the results obtained during the ensemble

application are interpreted. In most cases, for their qualitative

interpretation, it is necessary to contact specialists in the field of

medicine.

For automatic analysis and visualization, ready-made

analytical platform Deductor Studio was used, and, for

preprocessing and data cleaning, Loginom software products

developed by BaseGroupe Labs [18, 19] were used.

Thanks to the preprocessing of the initial sample, it is

possible either to level out completely or to minimize the errors

related to the human factor, which makes it possible to speak

about sufficient correctness of the medical research.

Unfortunately, small semi-structured samples cannot be called

sufficiently representative. However, the identified patterns may

be of interest to experts or for further verification.

4. Proposed Practical Solution

Laboratory test analysis has been carried out in workers

whose occupation is related to harmful factors (Group I) and a

control group (Group II).

Two small samples have been formed: the first sample

includes 100 records; the second one includes 207 records for

each of the scores taken into account for each group. A total of 9

scores of the general blood test were selected.

Data on the primary disease incidence in workers from

Groups I and II have been analyzed. The results of diagnoses

processing are presented.

The results of the study made it possible to identify the

diagnoses that most often occurred in Group I and Group II:

E78.0 – Pure hypercholesterolemia;

G90 – Disorder of autonomic nervous system, unspecified;

H52.4 – Presbyopia;

I10 – Essential [primary] hypertension.

The cross-table as an interactive tool for data representation

and analytical processing has allowed creating a pivot table

which represents data on quantitative composition of diseases for

each group.

The use of Kohonen self-organizing maps helped to identify

differences and patterns between the studied groups.

Examples of Kohonen self-organizing maps used are

presented in Fig. 2.

Fig.2. An example of data processing results using

Kohonen self-organizing maps

It should be noted that in both groups at least 24% of

employees were people aged 54 and over. In this regard, it has

been suggested that these diseases may be associated primarily

with age-related changes, for example, age-related decrease in

the accommodative ability of the eye associated with the natural

process of aging of the lens, a prolonged and persistent blood

pressure increase, blood cholesterol increase.

Also, this age group is characterized by significant positive

deviations from the norm in terms of hemoglobin, red blood

cells, and erythrocyte sedimentation rate.

The increase in red blood cells is especially noticeable in the

group of individuals whose work is associated with harmful

occupational factors. Erythrocytosis could be caused by external

factors.

Survey results may indicate the presence of pathological

processes in the body. Additional studies are needed to diagnose

the disease that has caused high content of red blood cells.

The methodology for analyzing small semi-structured

samples of biomedical data considered in the article can be used

with adequate efficiency as an integral part of the risk analysis of

diseases related to harmful occupational factors. Application of

the proposed method is demonstrated on specific actual data,

collection and consolidation of which has been carried out using

medical automated information system. This provides reliable

evaluation of harmful occupational factor effects on working

population’s health indicators.

The research results will help to define and evaluate

occupational risk factors that increase the likelihood of disease

development and to draft proposals for the prevention of harmful

effects of occupational factors on working citizens’ health.

5. Conclusion

The article discusses a methodology tested on specific actual

data, which allows pre-processing, cleaning and analysis of small

biomedical data samples.

As a result of the analysis, there have been revealed no

statistically significant difference in blood indices in individuals

from groups I and II.

Also, the study of the diagnoses of the primary disease

incidence in both groups has not revealed statistically significant

differences between the groups.

However, the established significant deviations from the

normal content of red blood cells and eosinophils in the blood of

individuals from Group I may indicate the presence of certain

pathological processes in the body, in particular, autoimmune

processes, identification of which requires additional research.

In the future work perspective, it is advisable to conduct a

research on these clinical blood test results using the traditional

analysis of variance method regardless of normal intervals. This

will allow for comparing the results obtained. Hereupon, it will

be possible to draw final conclusions about the influence of

specific occupational factors on the development of pathological

processes.

The results obtained help to increase the efficiency of

detecting patterns in biomedical databases and are of interest for

the construction of intelligent systems designed to analyze and

assess human health.

6. Acknowledgements

The reported study was funded by RFBR, project number 19-

07-00844.

7. Литература

[1] Bruce McCormick (2014) Update in Anaesthesia.

World Federation of Societies of Anaesthesiologists. 466 p.

[2] Kohonen T. The Self-Organizing Map // Proceeding of

the IEEE. 1990. Vol. 78. P. 1464-1480.

[3] Manyika J., Chui M., Brown B. et al. (2011) Big Data:

The Next Frontier for Innovation, Competition, and Productivity

/ McKinsey Global Institute.

[4] Baranov A.A., Namazova-Baranova L.S., Smirnova

I.V. et al. Methods and Tools for Complex Intelligent Analysis

of Medical Data. Trudy ISA RAN. Vol. 65. 2.2015. pp. 81-93.

[5] Barriers and Prospects for Digital Transformation: Big

Data Management Issues in the Healthcare Industry [Online]. –

Available: http://www.medlinks.ru/article.php?sid=83028

(Accessed: July 23, 2019).

[6] Geger E.V., Fedorenko S.I. Information Support of

Decision-Making when Assessing the Risk for Occupational

Morbidity Based on Analysis of Binary Samples // Proceedings

of Southwest State University. Control, Computer Engineering,

Information Science. Medical Instruments Engineering, no. 2

(27). 2018. pp. 101-107.

[7] Healthcare will Show the Largest Increase in Data

Generation by 2025 [Online]. – Available:

http://apcmed.ru/news/news-all/zdravookhranenie-pokazhet-

naibolshiy-rost-v-generatsii-dannykh-k-2025-godu/ (Accessed:

July 23, 2019).

[8] Izmerov N.F., Actualization of Occupational

Morbidity Issues // Health Care of the Russian Federation

(Zdravookhraneniye Rossiyskoy Federatsii.), no. 2. 2013. pp. 14-

17.

[9] Ismailova L.N. Effective Management of Production

Risks. // Economy and Business: Theory and Practice. 2016. No.

5. pp. 77-79.

[10] Kostenko N.A. Working Conditions and Occupational

Morbidity as a Basis for Risk Management of Workers' Health:

abstract of cand. med. sci. diss. M., 2015. 21 p.

[11] International Classification of Diseases of the tenth

revision (ICD-10) [Online]. – Available: https://mkb-10.com

(Accessed: July 20, 2019).

[12] The Order of the Ministry of Health and Social

Development of the Russian Federation No. 417n dated

27/04/2012 “On the Approval of the List of Occupational

Diseases”. [Online]. – Available:

http://base.garant.ru/70177874/ (дата обращения 23.05.2019).

[13] The Order of the Ministry of Health of the Russian

Federation dated 12/04/2011 No. 302n (as amended on

06/02/2018) “On the Approval of Lists of Harmful and (or)

Hazardous Occupational Factors and Kinds of Work that Require

Mandatory Preliminary and Periodical Medical Examinations

(Surveys), and the Procedure for Conducting Mandatory

Preliminary and Periodical Medical Examinations (Surveys)”.

[Online]. – Available: http://base.garant.ru/12191202/

(Accessed: May 21, 2019).

[14] The Federal Law dated 27/07/2006 No. 152-FZ (as

amended on 29/07/2017) "On Personal Data". [Online]. –

Available: http://base.garant.ru/5635295/ (Accessed: May 21,

2019).

[15] Tsvetkova L.A., Cherchenko O.V. Big Data

Technology in Medicine and Healthcare in Russia and the World

// Information technologies for the Physician, 2016. No. 3. pp.

60-73.

[16] Tsygankova, I.A. Method of Intelligent Processing of

Biomedical Data [Text] / I.A. Tsygankova // Software Products

and Systems. – 2009. –no. 3. – pp. 120-123.

[17] Choporov O.N., Razinkin K.A. Optimization Model of

Choice of the Initial Plan of Control Actions for Medical

Information Systems / Control Systems and Information

Technology. 2011. Vol. 46. No. 4.1. P. 185-187.

[18] BaseGroup. Data Analysis Technologies [Online] //

Available: https://basegroup.ru/ (Accessed: May 20, 2019).

[19] Loginom [Online] // Available: https://loginom.ru/

(Accessed: May 20, 2019).